Fasil Gadjimuradov1,2, Thomas Benkert2, Marcel Dominik Nickel2, and Andreas Maier1
1Pattern Recognition Lab, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany, 2Magnetic Resonance, Siemens Healthcare GmbH, Erlangen, Germany
Synopsis
Partial Fourier (PF) acquisition allows to reduce TE
in single-shot echo-planar imaging in order to increase signal-to-noise ratio
(SNR) in diffusion-weighted imaging (DWI). However, when applying it to
motion-prone liver DWI, conventional PF reconstruction methods fail since they
rely on smoothness priors of the phase. This work proposes to use an unrolled
network architecture which aims to estimate a more appropriate regularization
by learned recurrent convolutions. It can be shown that reconstructions
produced by the network are superior in terms of quantitative measures as well
as qualitative impression compared to conventional methods which tend to
introduce artifacts.
Introduction
DWI is valuable for lesion detection and
classification and is thus frequently used in clinical protocols for liver MRI
[1,2]. However, the T2*-blurring of typically employed single-shot echo-planar
imaging (ssEPI) combined with strong diffusion encoding leads to inherently
SNR-starved images. PF-sampling along the phase-encoding direction is often employed
to shorten the TE and to increase the apparent SNR.
In order to reconstruct an image from an
asymmetrically sampled k-space, conventional PF methods try to exploit
Hermitian symmetry in frequency domain by assuming smoothness of the phase. Homodyne
[3] and Projection Onto Convex Sets (POCS) [4] utilize a low-resolution (LR)
phase prior which is estimated from the symmetrically sampled part around the
k-space center. However, when applying these methods to liver DWI – which is often
subject to rapid, motion-related phase variations – artifacts are introduced (Figure
1). The fact that currently no robust reconstruction method for PF-DWI of the
liver exists, prevents its use in clinical practice or limits it to weak PF
factors with relatively small SNR gains.
This work aims to enable the application of PF to liver DWI using an
iterative reconstruction which, similar to POCS, alternates between
regularizing the current estimate and enforcing data fidelity. However, instead
of explicitly relying on smoothness assumptions of the phase, this work
proposes to learn a more suitable regularizer from training data using a recurrent
convolutional network.Methods
Network architecture: The
employed network architecture is derived from unrolling a variational splitting
algorithm for a fixed number of iterations. Within each iteration the algorithm
alternates between gradient descent with respect to regularizer and data term, respectively,
whereas the former is replaced by a neural network. The Deep Recurrent
Partial Fourier Network (DRPF-Net) employed in this work was introduced in
[5] and uses convolutional gated recurrent units [6] which conserve a memory state
across iterations (Figure 2). This allows weight-sharing across iterations
while maintaining the flexibility to adapt to iteration-specific conditions. While
multiple repetitions of a slice were arranged into a 3-D stack in [5], they are
now represented as a typical input batch allowing to employ 2-D instead of 3-D
convolutions which are expensive with respect to both run-time and memory. In
order to still exploit correlations across the repetitions and additionally
ensure permutation-equivariance, the idea of Deep Sets [7] was employed by
including pooling operations along the batch dimension.
Data:
Liver
DWI (b-values: 50, 800 s/mm2) was acquired in 25 volunteers using a
prototypical ssEPI sequence on 1.5 and 3 T MR scanners (MAGNETOM, Siemens
Healthcare, Erlangen, Germany). Data was acquired without PF but with parallel
acceleration (2x). Coil-combined, complex-valued ground-truth data was
generated using a prototypically modified reconstruction pipeline. Data was
divided into training (19 volunteers), validation (2) and test (4) sets
comprising 665, 70 and 140 image slices, respectively. In addition to that, one
data set was acquired using prospective PF-sampling (5/8).
Training: During
training, complex images were retrospectively sub-sampled with a PF-factor of
5/8 and then fed to the network as a two-channel input representing real and
imaginary part, respectively. The training objective was to minimize a loss
function consisting of two terms: a pixel-wise L1-loss and a perceptual loss [8],
where the former was weighted twice as high. Optimization was
performed using Adam [9] with a learning rate of 10-4
for 200 epochs.
Evaluation: The
reconstructions produced by the DRPF-Net and
POCS (each using 5 iterations) were evaluated both quantitatively and
qualitatively on retrospectively sub-sampled data. For qualitative analysis,
results on the prospectively PF-sampled data set were assessed as well. All
evaluations were performed on the averaged DW images and for a PF factor of
5/8.Results & Discussion
The qualitative comparison for retrospectively
sub-sampled data is presented in Figure 3. While the zero-filled reconstruction
is significantly blurred, POCS alleviates blurring but introduces signal
fluctuations and artificial structures due to the underlying LR phase estimate.
In contrast, the DRPF-Net produces sharp reconstructions without noticeable
artifacts. This is substantiated by the results of the quantitative evaluation (Figure
4) as the DRPF-Net outperforms zero-filling and POCS significantly with respect
to peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) for both b-values.
Figure 3 additionally demonstrates that the location
in the left liver lobe which suffers from artifacts in the POCS reconstruction
coincides with strong phase variations which cannot be reconstructed accurately
as shown for a selected repetition. In contrast, the phase map produced by the
DRPF-Net is in higher agreement with the ground-truth indicating the method’s
capability to recover high-frequency phase information even at high degrees of
asymmetry.
The performance on prospectively sub-sampled data can also be validated
as Figure 5 demonstrates the network’s ability to reconstruct visually
appealing images with homogeneous liver signal and high resolution with respect
to small structures like vessels.Conclusion
This work demonstrates that PF-reconstruction of motion-affected
liver DWI is possible using an unrolled network architecture. Due to the
elimination of phase priors unsuitable for the given application and the
employment of learned recurrent convolutions, the proposed method produces
quantitatively and qualitatively convincing reconstructions. Realizing liver
DWI with significantly reduced echo-train length can enable increased SNR and TE-intensive
schemes such as bipolar or flow-compensated diffusion preparation, for example.Acknowledgements
No acknowledgement found.References
[1]
Mitchell, D. G., Bruix, J., Sherman,
M., & Sirlin, C. B. (2015). LI‐RADS (Liver Imaging Reporting and Data System): Summary, discussion, and
consensus of the LI‐RADS Management Working
Group and future directions. Hepatology, 61(3),
1056-1065.
[2]
Taouli, B., & Koh, D. M. (2010). Diffusion-weighted MR imaging of the liver. Radiology, 254(1),
47-66
[3]
Noll, D. C., Nishimura, D. G., &
Macovski, A. (1991). Homodyne detection in magnetic resonance imaging. IEEE transactions
on medical imaging, 10(2), 154-163.
[4]
Haacke, E. M., Lindskogj, E. D., & Lin, W.
(1991). A fast, iterative, partial-Fourier technique capable of local phase
recovery. Journal of Magnetic Resonance (1969), 92(1), 126-145.
[5]
Gadjimuradov, F., Benkert, T., Nickel, M.
D., & Maier, A. (2020). Deep Recurrent Partial Fourier
Reconstruction in Diffusion MRI. In International Workshop on Machine Learning
for Medical Image Reconstruction (pp. 38-47). Springer, Cham.
[6]
Ballas, N., Yao, L., Pal, C., & Courville,
A. (2015). Delving deeper into convolutional networks for learning video
representations. arXiv preprint arXiv:1511.06432.
[7]
Zaheer, M., Kottur, S., Ravanbakhsh, S.,
Poczos, B., Salakhutdinov, R. R., & Smola, A. J. (2017). Deep sets. Advances
in neural information processing systems, 30, 3391-3401.
[8]
Zhang, R., Isola, P., Efros, A. A., Shechtman,
E., & Wang, O. (2018). The unreasonable effectiveness of deep
features as a perceptual metric. In Proceedings of the IEEE conference on
computer vision and pattern recognition (pp. 586-595).
[9] Kingma, D. P., & Ba, J. (2014).
Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980.
[10] Maier, A. K., Syben, C., Stimpel, B., Würfl,
T., Hoffmann, M., Schebesch, F., ... & Christiansen, S.
(2019). Learning with known operators reduces maximum error bounds. Nature
machine intelligence, 1(8), 373-380.